I Ran Germany's New #1 Open Source AI Model... Kolibri 1 🇩🇪

claims it is one of the best models via loaded language — Loaded Language (45/100)

At 0:18

Calling it 'one of the best' without a single benchmark is just hype 📈

Why this score: The speaker uses hyperbolic descriptors like 'one of the best' and 'so good' to create unearned hype without citing specific performance metrics or comparisons.

Original quote: “we're checking out Germania Almania Dutlands. One of the best models coming out of that country and it's so good.”

Sources: The 10 Best Large Language Models (LLMs) in 2026, A scalable framework for evaluating multiple language models through cross-domain generation and hallucination detection | Scientific Reports, 10+ Large Language Model Examples

claims German origin despite Canadian acquisition via missing context — Missing Context (45/100)

At 0:25

Trying to dodge the acquisition by saying it 'still counts' is a reach 🚩

Why this score: The speaker acknowledges the Canadian buyout but immediately attempts to pivot back to its 'German' identity, glossing over how ownership changes the nature of a 'national' model.

Original quote: “They recently got bought out by a big Canadian company. It's still based in Germany, so it still counts.”

Sources: German Canadians — Grokipedia, German Traces in Canada - Goethe-Institut Canada, Resignations shake Canadian art gallery after it rejects Nan Goldin acquisition over antisemitism claims - Jewish Telegraphic Agency

Uses historical dominance to claim Germany is 'taking over the world' via AI — Loaded Language (45/100)

At 0:30

Using the Roman Empire and royal lineages to justify a tech takeover is peak drama 🎭

Why this score: The speaker uses massive historical generalizations to create a sense of inevitable dominance, connecting ancient conquest to modern AI space.

Original quote: “Germany, they are known currently for being good engineers, but back in the day, they used to be conquerors. They were one of the first people to take over the Western Roman Empire. They founded France, the British royal family, German. The current president of America, German.”

Sources: The making of AI society: AI futures frames in German political and media discourses, Germany’s AI Offensive: The Ultimate Play for Global Power, Germany's AI strategy: Closing the gap with the US and China

Claims the model isn't actually good at coding despite its marketing — Confidence Mismatch (45/100)

At 1:06

Contradicts the model's own claims with a 'trust me' vibe 💀

Why this score: The speaker acknowledges the model claims to be good at coding but immediately pivots to a personal, unverified assessment that it fails in English.

Original quote: “It does say it's good at coding, but I'm just going to cut it short. It's not that good at coding. At least when I prompt it in English, it's not that good at coding.”

Uses specific benchmark scores to claim dominance over competitors — No Frame (75/100)

At 1:36

Actually brings the receipts for once. The numbers are there 📈

Why this score: The speaker provides specific comparative metrics (75 vs 60s) and references a specific benchmark (SWE) to support the model's performance.

Original quote: “overall in English it's smashing out they are getting 75 whereas the other ones they are in the 60s quen 3.5 is the closest one and if we look at coding so for example SWE everyone loves a little bit of SWE on average it's 88.9 so that's the highest in their level”

Claims model solves math with massive token output via 'thinking' mode — Confidence Mismatch (45/100)

At 2:30

He's bragging about '12,000 tokens of inference' like it's a miracle instead of just brute-force computation 💀

Why this score: The speaker is treating high token consumption during a 'thinking' process as a success metric for intelligence, rather than just an expensive computational cost.

Original quote: “it got the wrong answer with thinking disabled. ... But when I gave it low low thinking, it produced 12,000 tokens of inference, but it got the right answer.”

Claims model is capable in Swift despite imperfect responses — Missing Context (45/100)

At 2:59

Admits it's 'not the best answer' but calls it a win anyway. That's some weak-ass validation 🚩

Why this score: He is hand-waving the quality of the code, dismissing 'not the best' to maintain a positive narrative about its utility.

Original quote: “So, it does know about vectors C++. Does it know about Swift? ... it does actually give the right response.”

Claims model fails specifically on WinRT tasks even with high reasoning — No Frame (75/100)

At 3:40

Actually being honest about where the model breaks. No tricks here, just a real-world failure 🛡️

Why this score: He is providing a specific negative use case, which adds credibility to his overall testing process by not pretending it's perfect.

Original quote: “So, if you're doing WinRT, this is not the model for you. And it gets it wrong even with thinking at high.”

Claims model passed a logic test about distributing oranges — No Frame (75/100)

At 4:17

Passed a basic logic test. Not breaking any laws of physics or logic here ⚖️

Why this score: This is a straightforward demonstration of the model's ability to process context and avoid absurdities in a logic puzzle.

Original quote: “He knows you shouldn't cut the knife with the knife. So, each child gets two oranges.”

Claims 'thinking at low' fixes logic errors — a massive confidence leap. — Confidence Mismatch (45/100)

At 4:33

He's acting like a 'low' setting is some kind of magic fix for logic, but it's just a different way to fail 💀

Why this score: The speaker pivots from admitting failure to claiming success with 'thinking at low' without showing the actual logic that makes it work. It feels like he's trying to sell a broken tool by highlighting its only functional setting.

Original quote: “wrong answer with thinking disabled. Now, thinking at low, it does smash it out and get the right answer with drive.”

Admits the model fails even with thinking — a quick pivot in certainty. — Volume Game (45/100)

At 4:56

He builds up the hype only to immediately crash it. That's not a demo, that's a rollercoaster 🎢

Why this score: He makes an aggressive claim about the model's ability to solve logic, then immediately follows it with a failure that invalidates his previous confidence. It's loud-claim, quiet-retraction.

Original quote: “And even with thinking, what is this? Thinking at low, it gets the wrong one. So, I kind of take back what I just said.”

Claims Canadian acquisition means future improvement — a speculative leap. — Missing Context (45/100)

At 5:18

Claiming a buyout equals 'improvement' is pure speculation without knowing the actual roadmap 🇨🇦

Why this score: He uses a corporate event (the acquisition) to predict quality improvements. He's missing the context of why or how a change in ownership leads to better code/models.

Original quote: “they recently bought bought out by Canada, so they probably will be improving in the situation.”

Claims temperature zero is 'most accurate' — a technical simplification. — Loaded Language (45/100)

At 5:36

Calling temperature zero 'most accurate' is a massive simplification of how probability works 🎲

Why this score: He uses the word 'accurate' to describe a deterministic output (temp 0), which is technically just 'consistent.' He's using loaded terminology to make a technical setting sound like a quality guarantee.

Original quote: “it gets the most accurate responses you don't play around with the number random number generator”

Claims model lacks cultural nuance via anecdotal failure 🎭 — Personal Story (60/100)

At 7:34

Using a specific UK grime reference to prove it's 'unintelligent' is just a vibe check 💀

Why this score: He's using a single niche cultural failure as a proxy for the model's entire reasoning capability. It's an anecdote, not data.

Original quote: “So, it thinks it could be from a video game or a movie or an internet meme, it doesn't know about the the grime scene, unfortunately.”

Claims model is 'destroying' competitors with 88.9 score 🚩 — Loaded Language (45/100)

At 8:04

'Destroying' is a hell of a way to describe a benchmark lead 💀

Why this score: He's using aggressive, combat-heavy language ('destroying') to describe statistical performance differences. It masks how narrow the margin actually is.

Original quote: “So, SWE, we can see that on average it is 88.9. So, it's destroying GLM 4.7 Flash.”

Claims it competes with massive models like GPT-4 class via 88% score — Confidence Mismatch (45/100)

At 8:30

Throwing out a single '88' score to claim it's in the big leagues. That's some reckless hype 📈

Why this score: The speaker uses a single metric (88) to equate it with massive 120B+ parameter models without explaining the benchmark or context. It's a leap of faith based on one number.

Original quote: “going to just see this one more time cuz I'm just a bit surprised. So, English 88. So, it's up there with uh GPTOSs, the 120B, the big guy. It's up there with Mistro.”

Claims it's better than GLM 4.5 Air — Missing Context (45/100)

At 8:43

Claims it's 'beating' another model but doesn't say on what metric. Vague victory 🚩

Why this score: Saying one model is 'better' than another without specifying the benchmark (coding, reasoning, etc.) makes it a hollow victory. It's just noise.

Original quote: “And it's um going up against GLM 4.5 air. That was that was a breath of fresh air back in the day. It's even beating that. Better than GLM Air.”

Claims high efficiency based on 77GB memory usage — Loaded Language (45/100)

At 9:10

Calling it 'very, very efficient' based on memory usage alone is a stretch 💀

Why this score: Efficiency in AI usually involves performance-to-compute ratios. Labeling it 'very efficient' just because of a memory footprint is using loaded adjectives to mask technical nuance.

Original quote: “I'm running in debug. It's only going to get faster using 77 jibs of memory. So, it's very, very efficient.”

Claims more tokens equals better results with 'insane' hype — Emotional Button (45/100)

At 9:30

Using 'insane' to describe a token increase. That's just hype-man energy 🎭

Why this score: The speaker uses high-intensity emotional language ('insane', 'yozers') to describe standard technical scaling of context/tokens, turning a process into a spectacle.

Original quote: “So, with medium thinking less tokens and you get better generation. That's insane. So, high thinking we jumped from 3,500 tokens to 10,000 tokens.”

claims it is 'very very impressive' despite some minor issues — Confidence Mismatch (45/100)

At 10:30

Calling it 'very very impressive' while immediately pivoting to how he hoped it would do better 🎭

Why this score: He's using high-intensity adjectives like 'very very impressive' to mask the fact that he's already finding flaws in its output.

Original quote: “this is the Plotly version. So it's good they gave us two versions. Now this is the sun and I'm guessing these are planets.”

claims the model has 'beef' due to massive token output — No Frame (75/100)

At 10:56

Calling out the massive token count as a potential flaw—actually just calling it like he sees it 💀

Why this score: He's identifying a specific technical quirk (high token generation) and labeling it as 'beef' without trying to hide the inefficiency.

Original quote: “And with that, it produced 52,000 tokens. So there is some beef with this model. 52,000 tokens.”

claims the model switches from English to German mid-task — No Frame (75/100)

At 11:35

Reporting the weird language-switching behavior without any dramatic flair 🚩

Why this score: He is describing a specific, observable behavior of the model's output during code generation.

Original quote: “it started off in English. It just went up started doing some English and then when it came to the actual coding, it started printing out the code in German.”

claims the model's code results in runtime errors — No Frame (75/100)

At 12:09

Admitting the code it spits out fails to run. No lies detected here 💀

Why this score: He is documenting the failure of the generated code in real-time, providing no false sense of success.

Original quote: “it's using Pygame, but unfortunately that has a runtime error. So that is situations happening here.”

Dismissing current performance as just a 'version one' to dodge criticism — Missing Context (45/100)

At 13:07

Using the 'v1' excuse to handwave away current flaws 🚩

Why this score: He's using the novelty of a first release to lower the bar for quality. It's a classic way to avoid being held accountable for current bugs.

Original quote: “But again, this is a version one. So, they're only going to get better here.”

Warning about medical knowledge with a 'use at your own risk' disclaimer — Confidence Mismatch (45/100)

At 13:51

Giving a heavy warning about medical accuracy without actual data 💀

Why this score: He's pivoting from showing weird anatomy to a cautious warning, but he doesn't actually cite why it fails medical benchmarks—he just warns you to be careful.

Original quote: “I don't think it's that good for medical knowledge. Do not prompt it for that. At your own risks.”

Expressing shock at the massive drop in token output — No Frame (75/100)

At 14:06

Just calling out the erratic token behavior as it happens. No tricks here ⚖️

Why this score: He's reacting to the actual observed data of the model's performance dropping significantly during different prompts. It's a straight observation.

Original quote: “It's crazy how it went like 50,000 tokens, 7,000 tokens, 3,000 tokens.”

The model produces chaotic outputs that feel like a fever dream 🌀 — Just Vibes (50/100)

At 14:36

He's just reacting to the madness of the generation. Pure chaos energy 💀

Why this score: The speaker is reacting to the unpredictable nature of the model's output in real-time, using hyperbole rather than a technical claim.

Original quote: “I don't know what I don't know if this is Red Dead Redemption or Superman, but we're gone. We're gone, baby. Gone.”

The model's output is inconsistent and disappears mid-generation 📉 — Confidence Mismatch (45/100)

At 15:01

Complaining about the model's stability like it's a personality flaw 🎭

Why this score: He is expressing frustration with the model's ability to maintain a consistent output, but it's more of an observation of its current state than a hard metric.

Original quote: “Unfortunately, never to be seen again. Be careful. Be careful. Got to keep them in sight.”

Claims the model struggles with complex coding tasks 💻 — No Frame (75/100)

At 15:36

Admitting the coding performance is shaky after seeing it fail 💀

Why this score: The speaker provides a direct observation of the model failing to complete a 24k token coding task, which aligns with his earlier setup about its limitations.

Original quote: “24,000 tokens for a 3D platform game. We have game overed straight away. I I did say coding it wasn't the best.”

Suggests the model is only useful for basic business logic 💼 — Missing Context (45/100)

At 16:01

Limiting its utility to 'business logic' without explaining why 🚩

Why this score: He is attempting to find a niche for the model after seeing it fail at complex simulations, but he doesn't explain what 'business logic' actually means in this context.

Original quote: “So maybe with business kind of like logic applications, it's usable. Definitely look at the sun.”

gives a round of applause for basic functionality 👏 — Confidence Mismatch (45/100)

At 17:09

Giving a standing ovation for just changing a cube's color? That's some cheap hype 🤡

Why this score: He's celebrating extremely basic manipulation as a massive win for the model, masking its actual complexity behind enthusiasm.

Original quote: “it's got to get some points for that. I give it the round of applause. Calibri one smashed it out.”